Lect24 - Announcement Homework available on web site Nave...

This preview shows pages 1–8. Sign up to view the full content.

Announcement Homework available on web site Naïve Bayes and Decision Trees Relevant Talk tomorrow (Friday) Prof. Jeff Siskind Purdue Embodied Intelligence 3405 SC, 2PM 1

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Perceptron Decision Boundary Compare weighted sum of inputs to a threshold Without loss of generality set x 0 = -1 then w 0 is This defines a decision surface Which is the equation of a hyperplane n i i i x w 1 0 0 n i i i x w 0 0 x w n i i i x w 2
Perceptron Learning (Widrow-Hoff or delta rule) w x + Percep w + - percep w ( x ) assigns + or 1 if w x > 0 (vector dot product) else it assigns or 0 err = label( x ) percep w ( x ) 0: correct -1: false pos 1: false neg Here, false neg: w x < 0 but it should be > 0 loss = distance from boundary = - err w x Want to adjust w i ’s to reduce this loss Loss fcn gradient is direction of greatest increase in loss with w Want the opposite: step w in direction - w loss 3

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Perceptron Learning (Widrow-Hoff or delta rule) w x + Percep w + - Loss function = - err w x Want the opposite: step w in direction - w loss What is w loss? View - err w x as a function of w w (- err w x ) = - err x So - w (- err w x ) = err x Update w according to: w = err x where is a learning rate 4
Perceptron Learning (Widrow-Hoff or delta rule) w x + Percep w + - Choose a learning rate Compute w = err x Add w to w w 5

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
w x + Percep w + - Choose a learning rate ; initialize w arbitrarily (small works best) Compute w = err x Add w to w Repeatedly cycle through training examples Learn (always and only) on errors w New perceptron rotates to reduce error If x were a false positive… Perceptron Learning (Widrow-Hoff or delta rule) 6
If the points are linearly separable, the algorithm a) will halt b) will find a separator (The celebrated Perceptron Convergence Theorem) Choosing wisely will speed convergence If the points are not linearly separable, the algorithm may not halt. WHY?

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
This is the end of the preview. Sign up to access the rest of the document.

Lect24 - Announcement Homework available on web site Nave...

This preview shows document pages 1 - 8. Sign up to view the full document.

View Full Document
Ask a homework question - tutors are online